A consensus is forming in the agent world: with the same frozen model, swapping the surrounding "harness" — the prompts, control flow, tooling, and context management — can change capability dramatically. So a growing number of teams let agents modify their own harnesses: iteratively proposing edits, evaluating on benchmarks, and keeping the winners. This is recursive self-improvement (RSI) at the agent-system level. A Google team's paper released on September 21, RRSI, pours cold water on this route: a harness evolved this way may simply be memorizing the test.
The Disease: Harness Evolution Overfits
The paper starts with a diagnosis: evolving a harness around a frozen model delivers visible gains, but recursive evolution overfits — the harness memorizes the training tasks, posts large in-distribution gains, and those gains shrink or vanish on out-of-distribution (OOD) benchmarks. It is structurally the same problem as a deep net memorizing its training set, except the object being memorized with is prompts and control flow instead of weights.
Two-Sided Regularization: Cap the Proposals, Prune the Rest
RRSI embeds regularization principles into both ends of the evolution loop. On the proposal side, the proposer operates with a temporally annealed budget that caps how many independent edits one candidate may bundle; it is conditioned on the full edit history so falsified hypotheses are not redrawn, and stalled runs get redirected toward components never exercised. On the selection side, the selector is equipped with a critic and a pruner: the critic screens benchmark-specific proposals before evaluation, while the pruner removes changes that are too small, too expensive, or no longer useful. The GitHub README adds two more details: a noise-adjusted floor blocks gains within evaluation variance, and a cost rule requires any added inference tokens to be paid for by measured gain.
The Numbers: +4.7 OOD, 30% Fewer Tokens
Across eight benchmarks spanning coding, agentic workspace, and engineering design tasks, RRSI gains up to 14.1 points on the split it evolves against and up to 4.7 points on the five out-of-distribution benchmarks, while producing a harness that runs on 30% fewer policy tokens than unregularized evolution — more general and cheaper. The same loop drives three instances: a terminal agent (Terminal-Bench 2.1), a document-work agent (Harvey LAB), and an engineering-design agent (EngDesign). The paper hit #1 on Hugging Face's daily papers chart on release day. Code and project page are live: repository google-research/rrsi, project page regularized-rsi.com.
So What
The value of this paper is not the score gains themselves but turning an engineering intuition into a reproducible methodology: harness optimization is a search, and every search needs regularization — an old idea borrowed back from deep learning, now reactivated at the agent-system level. Look at the neighboring work: DeepSeek open-sourced a pluggable harness plugin stack, NVIDIA's SoL-Pi cuts tokens with an external harness, and HarnessDev has LLMs build their own harnesses as an exam. Everyone assumes the harness deserves systematic optimization; what RRSI adds is the missing link — the optimization process itself can overfit, and must be defended against. For teams doing agent engineering, this is a direct warning: if your agent tuning keeps climbing on one fixed eval set, suspect memorization before you believe it got smarter.
References: paper arXiv:2609.24972 (https://arxiv.org/abs/2609.24972); code https://github.com/google-research/rrsi; project page https://regularized-rsi.com/